Our Products
Panda Agent Speech-to-Text System
Speech Recognition / Offline AI
Speech RecognitionOffline AIEdge ComputingVAD Segmentation
Product Overview
Based on the FasterWhisper Small model, it provides fully offline Chinese speech-to-text capability. Built-in VAD endpoint detection for automatic sentence segmentation, no internet required, supports CPU/CUDA inference, suitable for quadruped robot command parsing, industrial equipment voice control, intelligent security voice recording, and other scenarios.
Key Features
| Speech-to-Text | Real-time conversion of microphone input to Chinese text, supporting continuous recognition |
| VAD Segmentation | Silence threshold 800ms automatically detects speech boundaries, no manual trigger needed |
| Fully Offline | Local inference, zero network dependency, data stays on edge device |
| int8 Quantization | Reduces computation, improves CPU utilization efficiency |
| Multi-core Parallelism | Automatically utilizes multiple logical CPU cores |
| USB Microphone | Supports standard USB audio devices, 16kHz mono PCM capture |
| Result Output | Text results output to application layer in real time, supports JSONL logging |
| CUDA Acceleration | Switchable CUDA inference (Jetson Orin), significantly reduced latency |
Technical Specifications
| STT Model | FasterWhisper Small (CTranslate2 int8 quantized) |
| Language Support | Chinese (Mandarin), extensible to multilingual |
| Sampling Parameters | 16000Hz mono int16 PCM |
| VAD Threshold | Silence 800ms triggers segmentation |
| Runtime Memory | Resident ~405MiB, peak ~544MiB |
| Runtime Environment | Python 3.8+, faster-whisper, Ubuntu 20.04 |
Typical Applications
Quadruped Robot Voice CommandsIndustrial Equipment Voice ControlSecurity Voice RecordingSmart Q&A FrontendVoice Memos